Skip to content
Frontline AI Intelligence
Search intelligence中文

Evidence brief

DeepSeek: More memory is not always better for agents: evaluation of eight models shows dosage should be calibrated by capability

Agent memory is not a feature to switch on casually; its dose must be calibrated to model capability. Strong models are better suited to injecting a full set of guides, with DeepSeek-V3.2 (671B MoE) improving task completion by +9.5 percentage points. Weaker models perform best with curated retrieval, with gpt-oss-120b (117B MoE) improving by +16.1pp while adding only +5% tokens. This method requires no weight updates or manual annotation; it works by distilling guides from an agent’s past trajectories and injecting them at inference time.

Published
Updated
Editorial
Frontline Lab
Source
Hugging Face：Blog（RSS）
Source author
Hugging Face：Blog（RSS）
Related topics
0
Collected
2026-08-19

Frontline Lab summary and source

Editorial summary

Agent memory is not a feature to switch on casually; its dose must be calibrated to model capability. Strong models are better suited to injecting a full set of guides, with DeepSeek-V3.2 (671B MoE) improving task completion by +9.5 percentage points. Weaker models perform best with curated retrieval, with gpt-oss-120b (117B MoE) improving by +16.1pp while adding only +5% tokens. This method requires no weight updates or manual annotation; it works by distilling guides from an agent’s past trajectories and injecting them at inference time.

This brief preserves the original source so the summary and editorial context can be checked independently.

Source attributionHugging Face：Blog（RSS） · Hugging Face：Blog（RSS）

Open the original source

Related published evidence

Relationships are derived from shared topics, entities, categories, tags, and community context; every result remains independently source-linked.

OpenAI: official site updates (RSS · excluding enterprise/customer cases)

OpenAI slows model scaling due to critical cyber capability threshold

Due to the OpenAI-Hugging Face incident and the possibility that the Astra model may have reached a critical cybersecurity capability threshold, OpenAI temporarily slowed model scaling, paused reinforcement learning training for its latest deployed model for two weeks, and put its largest frontier RL run on hold.

Why it mattersThis measure shows that the pace of frontier model training is being constrained by cybersecurity capability evaluations.

Original source
X

How DeepSeek Harness desktop wrapper preinstalls features

The poster says the open-source DeepSeek Harness desktop client has been popular recently, and believes the DSH philosophy should be minimalism and everything as a plugin, questioning why some desktop wrappers hard-code features into the shell.

Why it mattersThis view points to the plugin boundaries and product form choices of desktop Agent tools.

Original source